Complex reasoning on GSM8K (test)
87.3AccuracyFull Repetition
Evaluation Results
| Method | Links | |
|---|---|---|
| Full RepetitionAvg KV Cache=471.1, Avg Prefill FLOPs=3.140 T, Backbone=Qwen 2.5-3B2026.07 | 87.3 | |
| Naïve SummaryAvg KV Cache=277.4, Avg Prefill FLOPs=3.676 T, Backbone=Qwen 2.5-3B, Repetition Type=Appending summary2026.07 | 85.5 | |
| PARTREPAvg KV Cache=280.4, Avg Prefill FLOPs=2.481 T, Backbone=Qwen 2.5-3B, τ=0.152026.07 | 85.5 | |
| No RepetitionAvg KV Cache=235.5, Avg Prefill FLOPs=1.549 T, Backbone=Qwen 2.5-3B2026.07 | 84.2 | |
| LLMLinguaAvg KV Cache=277.4, Avg Prefill FLOPs=2.105 T, Backbone=Qwen 2.5-3B, Repetition Type=Appending summary2026.07 | 84 | |
| Echo EvictionAvg KV Cache=235.5, Avg Prefill FLOPs=3.152 T, Backbone=Qwen 2.5-3B, Repetition Type=Compressing full repetition2026.07 | 76.6 | |
| H2O EvictionAvg KV Cache=270.8, Avg Prefill FLOPs=3.162 T, Backbone=Qwen 2.5-3B, Repetition Type=Compressing full repetition2026.07 | 72.9 | |
| LLMLingua Comp.Avg KV Cache=78.6, Avg Prefill FLOPs=1.130 T, Backbone=Qwen 2.5-3B, Repetition Type=Compressing full repetition2026.07 | 12.2 |